Smart power grid dispatching optimization method based on adaptive evolution control

Abstract

The invention provides a smart power grid dispatching optimization method based on adaptive evolution control, which comprises the following steps: performing power network graph modeling and standardization processing on a power network to generate a standardized graph structure; constructing a Markov decision framework of power grid dispatching based on the standardized graph structure; performing interactive training under a Markov decision framework to generate a preliminary scheduling strategy meeting power balance constraints; taking the preliminary scheduling strategy as an initial population, and generating a global optimization scheduling scheme through fitness function evaluation and multi-mode crossover mutation operation; and deploying a global optimization scheduling scheme to an actual power grid, and dynamically updating strategy network parameters based on node loads, power generation output and frequency data acquired in real time. According to the method, the power grid dispatching strategy can be dynamically adjusted in real time, consumption and dispatching of new energy are optimized, and the adaptive capacity of the power grid to different loads and power generation fluctuations is improved.

Description

Technical Field

[0001] The present invention relates to the field of novel power systems, and in particular to a smart grid dispatch optimization method based on adaptive evolutionary control. Background Art

[0002] The new power system is an important part of the new energy system. The main features of the new power system include a high proportion of renewable energy access, widespread distribution of distributed energy resources, and enhanced interaction between sources, grids, loads and storage. With the rapid development of new energy sources such as wind power and photovoltaics, my country's power grid structure has undergone profound changes, and the operating characteristics of the power system have become more complex, placing higher demands on grid balance. Existing power grid dispatch optimization methods mainly rely on mathematical modeling of the electricity consumption behavior of power generation equipment and users, and dispatching through optimization algorithms. However, these methods have many shortcomings when facing complex power system dispatch problems:

[0003] (1) Traditional optimization methods have poor adaptability. Traditional grid dispatching methods mostly rely on classical mathematical modeling and optimization techniques such as linear programming. These methods usually optimize the dispatching of generators and load distribution based on static assumptions, ignoring the dynamic interaction between different devices in grid operation. Especially when facing the large-scale access of renewable energy sources such as wind power and photovoltaics, the load volatility and uncertainty of the grid make it difficult for traditional optimization methods to quickly adapt to changes in the system, which may lead to grid imbalance and even system failure. Such methods have poor adaptability and cannot effectively cope with the dynamic changes in the grid caused by new energy.

[0004] (2) Insufficient model accuracy and reliability. Traditional scheduling optimization methods often rely on simplified mathematical models to simply model the power generation characteristics of equipment and the electricity consumption behavior of users. Most methods ignore the specificity between equipment and users and cannot fully capture the complex nonlinear relationships and spatiotemporal characteristics in the data. This leads to poor prediction accuracy and reliability of these models in practical applications. For example, the uncertainty and volatility of wind and photovoltaic power generation are not effectively modeled in traditional models, making the results of power grid scheduling unable to meet the high-precision scheduling requirements. As the power grid system becomes increasingly complex, the traditional methods are unable to model the data.

[0005] (3) Traditional methods lack the ability to process multi-source heterogeneous data. With the access of distributed energy resources such as renewable energy, energy storage equipment and electric vehicles to the power grid, the types and scale of data have exploded. However, traditional power grid dispatching methods have many problems in dealing with multi-source heterogeneous data. Most existing methods rely on simple historical data analysis or time series prediction, which cannot effectively integrate information from different data sources and fail to consider the relationship between data. This makes these methods lack the necessary flexibility and accuracy when processing large-scale, multi-dimensional data in the power grid, resulting in poor optimization of dispatching schemes.

[0006] (4) The application of existing reinforcement learning algorithms faces bottlenecks. In order to overcome the shortcomings of traditional optimization methods, reinforcement learning (RL) has begun to be applied to power grid scheduling problems in recent years. Reinforcement learning can learn optimal strategies through interaction with the environment and has strong adaptive capabilities. Especially in dealing with dynamic power grid scheduling, it can adjust decisions in real time. However, the training process of reinforcement learning requires a large amount of sample data, which may lead to low training efficiency in practical applications and make it difficult to meet the real-time requirements of power grid scheduling. At the same time, the explosion problem of state and action space also makes the reinforcement learning algorithm converge slowly in the exploration decision-making process.

[0007] With the large-scale integration of renewable energy sources such as wind and photovoltaics, power system scheduling faces increasingly complex challenges. Traditional grid scheduling methods are unable to effectively address the uncertainty and volatility of these energy sources, making it difficult to maintain grid balance and potentially leading to power shortages or over-scheduling. Furthermore, with the introduction of diverse distributed energy devices and multi-source heterogeneous data in the power grid, traditional scheduling methods have significant shortcomings in processing and integrating this complex data. Summary of the Invention

[0008] This paper proposes a smart grid dispatch optimization method based on adaptive evolutionary control. By combining reinforcement learning with genetic algorithms, it can dynamically adjust grid dispatch strategies in real time, optimize the absorption and dispatch of renewable energy, and improve the grid's adaptability to varying loads and power generation fluctuations. This approach achieves optimal allocation and utilization of power resources, ensuring user power needs while effectively reducing power losses and ensuring efficient, reliable, and stable grid operation.

[0009] A smart grid dispatch optimization method based on adaptive evolutionary control includes the following steps:

[0010] S1: Power network graph modeling and standardization. Wind power / photovoltaic equipment, energy storage equipment, and user loads in the power network are modeled as graph nodes. Edge connections are established between graph nodes through power transmission paths. Various dynamic data generated during the operation of the power grid are denoised, gap-filled, and normalized to generate a standardized graph structure containing device characteristic parameters.

[0011] S2: Markov decision process modeling, based on the standardized graph structure, defines the state space, action space, state transition probability, reward function and discount factor, and constructs a Markov decision framework for power grid scheduling;

[0012] S3: Reinforcement learning strategy generation, using the reinforcement learning algorithm to perform interactive training under the Markov decision framework to generate a preliminary scheduling strategy that meets the power balance constraint;

[0013] S4: Genetic algorithm global optimization, using the preliminary scheduling strategy as the initial population, generating a globally optimized scheduling solution through fitness function evaluation and multi-mode crossover and mutation operations;

[0014] S5: Adaptive feedback control deploys the globally optimized dispatching scheme to the actual power grid and dynamically updates the strategic network parameters based on real-time collected node load, power output and frequency data.

[0015] Furthermore, the step S1 specifically includes the following steps:

[0016] S1-1: Power Network Equipment Modeling

[0017] Wind power / photovoltaic equipment, energy storage equipment, and user loads are abstracted as nodes in a graph, and the key characteristics of the equipment are described as node attributes. Device nodes are connected based on power flow relationships to establish graph edges. The key characteristics include power generation capacity, storage capacity, and load variation characteristics. Edge attributes include transmission line capacity, power transmission capacity, and energy loss coefficient.

[0018] S1-2: Power system operation data collection and cleaning

[0019] Based on the S1-1 modeling framework, various dynamic data generated during the operation of the power grid are systematically collected. De-noising, outlier detection and elimination, missing data filling, and standardized conversion are applied to various dynamic data.

[0020] S1-3: Key feature extraction and data preprocessing

[0021] After completing the basic data cleaning, key characteristic variables are extracted from multi-source data based on the needs of power grid dispatch optimization, including historical load curves, wind and solar power generation forecasts, load fluctuation indicators, and source-load matching. The extracted key characteristic variables are normalized and dimensionalized to eliminate scale differences between different data sources and enhance the applicability and stability of the data in subsequent modeling, training, and inference stages.

[0022] Furthermore, the various types of dynamic data include equipment operating status, power generation output, load demand, meteorological conditions, and source-load scheduling direction.

[0023] Furthermore, the step S2 specifically includes the following steps:

[0024] The grid dispatch problem is modeled as a Markov decision process, which is defined as the following five-tuple:

[0025]

[0026] S2-1: Define the state space

[0027] The state space is used to describe the operating state of the power grid at a certain moment. It is defined based on the operating characteristics of the power grid. The state space contains various variables of the power grid, including wind power generation, photovoltaic power generation, the status of energy storage equipment, user load demand, grid frequency, and voltage. Each state represents the operating state of the power grid at a certain moment.

[0028] The state space formula is:

[0029] P t : Real-time power generation vector of each distributed power generation device;

[0030] Q t : The current charge and discharge state vector of the energy storage device;

[0031] D t : Real-time load demand vector of each user;

[0032] E t : The current topology of the power grid and the power flow status between nodes;

[0033] S2-2: Defining the action space

[0034] The action space is used to describe scheduling decisions. The action space is defined to represent the scheduling decisions that the grid can take in each state. The action space includes equipment startup and shutdown, generator output adjustment, energy storage device charging and discharging decisions, and load-side response control. Each action corresponds to a possible scheduling solution, which is used to guide the grid state toward the desired goal.

[0035] The action space formula is:

[0036] C t : charge and discharge control strategy vector of energy storage equipment;

[0037] F t : load demand adjustment strategy vector;

[0038] S2-3: Define state transition probability (P)

[0039] Based on the definition of state and action, the state transition probability is characterized, that is, the probability distribution of the power grid transferring from the current state to the next state under the action of a specific action, and the probability distribution of the power grid transferring from the current state s t When performing action a t Then transfer to the next state s t+1 possibility;

[0040] The state transition formula is: P(S t+1 |s t ,a t )

[0041] s t+1 It is determined by the grid dynamics and control behavior and is estimated by simulating the grid dynamics equations;

[0042] S2-4: Define the reward function (R)

[0043] Design a reward function to evaluate strategies based on the effectiveness of grid dispatch. The reward function comprehensively considers energy efficiency, power loss, minimization of dispatch costs, grid load balance, improvement of renewable energy absorption rate, and system operation stability, thereby guiding the model to gradually form a dispatch strategy that meets the overall optimization goal during the learning process.

[0044] The design of the reward function includes:

[0045] Grid balancing objective: to ensure the balance between power generation and load, defined as R balance =-|P t -D t |;

[0046] Scheduling cost objective: Minimize the cost of power generation and the cost of charging and discharging energy storage equipment, defined as R cost =-∑ i (c i P i,t +d i Q i,t ), where c i and d i is the cost coefficient of power generation and energy storage;

[0047] Network stability goal: reduce power flow fluctuations and line losses, defined as R stability =-∑ j∈lines L j , where L j is the power loss of the jth line;

[0048] Total reward function: R(s t ,a t )=α1R balance +α2R cost +α3R stability , where α1, α2, α3 are weight parameters;

[0049] S2-5: Define the discount factor (γ)

[0050] Set an appropriate discount factor to balance immediate rewards with long-term returns. By properly setting the discount factor value, the model is encouraged to focus not only on current scheduling benefits but also on the overall optimization effect of future system operations, thereby enhancing the strategy's foresight and robustness.

[0051] Steps S2-1 to S2-5 describe the power dispatch problem as a Markov decision process, whose solution goal is to find the optimal strategy π through reinforcement learning method. * , so that the cumulative reward is maximized:

[0052]

[0053] Furthermore, the step S3 specifically includes the following steps:

[0054] S3-1: Model initialization

[0055] Based on the definition of the Markov decision process, the reinforcement learning model is initialized, specifically including: constructing a fixed state space, action space, and reward function; using a parameterized neural network as an approximator of the policy function and the value function respectively; initializing the parameters of the neural network; the neural network adopts a dual network structure, corresponding to the policy network and the value network respectively, using the ReLU function as the activation function, and using the Adam optimizer as the optimizer;

[0056] S3-2: Environment Interaction and Strategy Learning

[0057] Construct a power grid simulation environment, simulate the power grid operation process based on the power flow calculation method, and dynamically generate time series data of clean energy power generation, user load demand, and energy storage device charging and discharging behavior; at each discrete moment t, the reinforcement learning model is used according to the current state s t And the policy network output π, determine the unique action a t After executing this action, the power grid environment is updated to the next state s according to the flow calculationt+1 , and returns the immediate reward R t ;

[0058] During the interaction process, the proximal policy optimization algorithm is used to synchronously update the policy network and the value network. At the same time, the experience replay mechanism is introduced to perform batch sampling training on historical interaction samples to improve the independence of training data and the generalization ability of the strategy.

[0059] S3-3: Strategy Evaluation and Hyperparameter Optimization

[0060] During the reinforcement learning training process, the dispatching strategy output by the model is evaluated at a fixed period. The evaluation indicators include: the balance error between total power generation and load demand, the trend of cumulative reward changes, and the stability of the power flow of the power grid, which includes voltage offset and frequency fluctuation amplitude.

[0061] Based on the evaluation results, the training hyperparameters are dynamically adjusted, including updating the exploration rate ∈ according to the preset linear decay strategy and adaptively adjusting the PPO clipping coefficient according to the reward stability to ensure the stability and convergence of the strategy optimization process. The update formula is:

[0062] ∈=max(∈ min ,∈ max ·e -k·t )

[0063] Among them, ∈ min is the minimum exploration rate, k is the decay factor;

[0064] S3-4: Preliminary scheduling strategy generation

[0065] After the reinforcement learning training reaches the preset stopping criteria, the final trained policy network is extracted as the preliminary scheduling strategy. The stopping criteria are that the cumulative reward reaches the set threshold and the reward fluctuation is less than the set range for multiple consecutive evaluation cycles.

[0066] Furthermore, the step S4 specifically includes the following steps:

[0067] S4-1: Generate initial population based on preliminary scheduling plan

[0068] (1) Definition of population individuals

[0069] Each population individual represents a complete set of grid dispatch plans, including the specific power output instructions of each power generation equipment and energy storage equipment at each dispatch time. The individual encoding adopts real number encoding, with a fixed encoding structure and a length proportional to the number of dispatchable devices in the grid and the dispatch duration. The population individual formula is defined as:

[0070] x i ={G i ,Ci ,F i}

[0071] in:

[0072] G i : Output adjustment strategy of power generation equipment;

[0073] C i : Charging and discharging power distribution of energy storage equipment;

[0074] F i : Load side response adjustment strategy;

[0075] (2) Generate the initial population

[0076] The preliminary scheduling strategy π generated by reinforcement learning RL , introducing a small random perturbation to generate the initial population, ensuring that the performance of the individuals in the initial population is close to the local optimum and has a certain degree of diversity; at the same time, the feasibility of the generated scheme is checked to ensure that it meets the grid operation constraints. Individuals that do not meet the constraints will be corrected or regenerated;

[0077]

[0078] Where N represents the population size, which is fixed at N = 100 individuals;

[0079] S4-2: Fitness evaluation and individual selection

[0080] Define the fitness function f(x i ), taking the overall performance of the grid dispatching scheme as the basis for fitness evaluation, and then evaluating the pros and cons of each dispatching scheme:

[0081] f(x i )=-(λ1·L balance +λ2·L stability +λ3·C cost )

[0082] L balance : Grid power balance error, which measures the degree of matching between power input and output;

[0083] L stability : Grid operation stability indicators, such as power flow fluctuations, voltage deviation, etc.;

[0084] C cost : The operating cost of the dispatch scheme, including energy storage losses and load adjustment costs;

[0085] λ1,λ2,λ3: weight coefficients, reflecting the relative importance of different optimization objectives;

[0086] An elite retention strategy combined with a roulette wheel selection mechanism is used to select outstanding individuals from the current population to enter the next generation based on the fitness value calculated by the fitness function, ensuring that excellent genes are passed on. The probability of an individual being selected is positively correlated with its fitness value.

[0087] S4-3: Genetic Manipulation and Evolution

[0088] (1) Crossover operation

[0089] Perform a crossover operation on the selected individuals to generate the next generation:

[0090] x new =α·x i +(1-α)·x j

[0091] where x i ,x j represents the parent individual, α is the crossover ratio, which is a random value;

[0092] For the cross operation of the grid dispatching scheme, three cross schemes are adopted: First, the generation dispatching part G in the dispatching scheme is directly i 、Energy storage dispatch part C i and demand response part F i Each of these three gene fragments is considered a gene fragment, and two parent individuals are randomly selected to cross the three gene fragments. Since power grid scheduling usually has time series characteristics, a scheduling cycle is divided into multiple time periods. From the two parent individuals, the scheduling strategies of different time periods are selected and combined into new individuals. Finally, since the power grid scheduling plan involves multiple regions, the scheduling strategies of the parent individuals are exchanged between different regions.

[0093] (2) Mutation operation

[0094] Perform genetic mutations on some individuals to enhance population diversity and avoid falling into local optimality:

[0095] x mut =x k +δ

[0096] Where δ is the random perturbation;

[0097] For the grid dispatching plan, the energy storage dispatching part C in the dispatching plan can be i and demand response part F i Perform random parameter perturbations to introduce larger random variations to lower-priority devices in the power grid, maintaining the dispatch stability of key devices.

[0098] Based on crossover and mutation operations, a new generation of population is generated: The population is optimized step by step through a loop. To ensure the quality of the solution, the solution with the highest fitness in the previous generation is directly retained in the population of the next generation.

[0099] S4-4: Output optimal strategy

[0100] When the change in the population fitness value is less than the set threshold, the genetic algorithm converges; eventually, the individual with the highest fitness is selected from the population, and the grid dispatching scheme corresponding to its code is extracted as the optimal dispatching strategy after genetic algorithm optimization:

[0101]

[0102] Furthermore, the step S5 specifically includes the following steps:

[0103] S5-1: Real-time feedback data collection and preprocessing

[0104] During the grid dispatch process under the guidance of the global optimization dispatching scheme obtained by S4 step optimization, the grid operation data is collected in real time, including: node load demand L i(t) , node power generation output G i(t) , remaining capacity of energy storage equipment S i(t) , grid system frequency f(t), line power transmission status P {ij}(t) ; Where i, j represent the node numbers in the network, and t is the current time;

[0105] The collected data is processed through standardization and anomaly detection to eliminate invalid or abnormal data;

[0106] S5-2: Feedback Learning and Incremental Update of Reinforcement Learning Model

[0107] Based on the real-time preprocessed data, the reinforcement learning model performs incremental updates. Specifically, the model learns according to the following rules:

[0108] Current status t It is composed of the collected real-time grid operation data;

[0109] Action a t Determined by the current scheduling policy;

[0110] Reward t Calculated based on actual grid operation results;

[0111] S5-3: Real-time evaluation and adaptive adjustment of strategy effects

[0112] During the feedback learning process, the updated scheduling policy is periodically evaluated. If the evaluation indicator exceeds the set threshold, the adaptive policy adjustment mechanism is triggered:

[0113] Temporarily increase the exploration action ratio;

[0114] Update the reward function parameter w i ;

[0115] Re-adjust the output distribution of the strategy network to enhance the model's adaptability to abnormal working conditions;

[0116] S5-4: Forming a continuous optimization closed loop

[0117] Through the continuous execution of steps S5-1 to S5-3, a closed-loop mechanism of data collection - feedback learning - strategy evaluation - adaptive adjustment is formed, so that the power grid dispatching strategy can dynamically adapt to external environmental disturbances such as load fluctuations, climate change and equipment status changes, and continuously improve the stability, reliability and economy of power grid operation; at the same time, the strategy optimized by the genetic algorithm is compared with the strategy initially generated by reinforcement learning to verify the optimization effect of the genetic algorithm.

[0118] Furthermore, the thresholds set in step S5-3 are that the load satisfaction rate is lower than 95% and the system frequency deviation exceeds ±0.1 Hz.

[0119] Compared with the prior art, the present invention has the following advantages:

[0120] (1) In terms of power grid modeling, wind power, photovoltaic power, energy storage equipment, and user loads are uniformly abstracted as graph structure nodes, and graph edges are constructed based on power flow relationships, which improves data consistency and standardization. Test results show that the node and edge attribute coverage rate reaches over 92%, and the modeling processing efficiency is improved by about 24% compared with the method based on manual maintenance of topological relationships, providing a solid foundation for scheduling modeling based on Markov decision processes.

[0121] (2) In terms of reinforcement learning, training grid dispatch strategies based on the PPO algorithm can accurately capture the complex relationship between renewable energy fluctuations and load changes, enabling real-time state perception and decision-making. Simulations show that the load satisfaction rate exceeds 87%, and the power generation forecast error is reduced by approximately 19% compared to traditional methods based on linear regression prediction, significantly enhancing the dynamic adaptability of the grid under the integration of renewable energy.

[0122] (3) In terms of scheduling strategy optimization, a genetic algorithm was introduced to conduct a global search based on the reinforcement learning preliminary strategy, effectively avoiding local optimality. Verification results show that compared with the simple reinforcement learning preliminary strategy (without subsequent optimization), the energy loss of the optimized strategy is reduced by about 17%, and the utilization rate of energy storage equipment is increased by about 12%, further improving the economy and stability of power grid scheduling. BRIEF DESCRIPTION OF THE DRAWINGS

[0123] Figure 1 This is a flow chart of a smart grid dispatch optimization method based on adaptive evolutionary control according to an embodiment of the present invention.

[0124] Figure 2 This is a Markov decision process modeling diagram of an embodiment of the present invention.

[0125] Figure 3 It is a closed-loop learning and optimization diagram of the scheduling strategy according to an embodiment of the present invention. DETAILED DESCRIPTION

[0126] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0127] See also Figure 1 The present invention provides a smart grid dispatch optimization method based on adaptive evolutionary control, which is used to improve the dispatch efficiency and stability of the power network under the large-scale access of renewable energy. By combining reinforcement learning and genetic algorithms, this method can achieve grid load balance and optimize the dispatch plan of renewable energy when facing complex grid dispatch problems. The method specifically includes the following steps:

[0128] S1: Modeling and standardization of power network graphs. Wind power / photovoltaic equipment, energy storage equipment, and user loads in the power network are modeled as graph nodes. Edge connections are established between graph nodes through power transmission paths. Various types of dynamic data generated during the operation of the collected power grid are denoised, gap-filled, and normalized to generate a standardized graph structure containing device characteristic parameters. Through modeling, complex multi-source power systems are unified into a standardized graph structure, facilitating subsequent mathematical modeling based on the Markov Decision Process (MDP) and the application of related algorithms. Subsequently, the scheduling optimization model will be trained based on the preprocessed data, and through parameter adjustment, scheduling strategies suitable for different operating scenarios will be learned and refined.

[0129] The step S1 specifically includes the following steps:

[0130] S1-1: Power Network Equipment Modeling

[0131] First, a graph structure modeling is performed for various types of equipment in the power system. Clean energy power generation equipment (such as wind power and photovoltaic power generation devices), energy storage equipment, and user loads are abstracted as nodes in the graph, and the key characteristics of the equipment (such as power generation capacity, storage capacity, and load variation characteristics) are described as node attributes. Based on the power flow relationship, the device nodes are connected to establish graph edges. The edge attributes include transmission line capacity, power transmission capacity, energy loss coefficient, etc. Through this process, the complex power network is mapped into a structured, fully attributed graph representation, providing the foundation for subsequent scheduling modeling.

[0132] S1-2: Power system operation data collection and cleaning

[0133] Based on this modeling framework, we systematically collect various dynamic data generated during grid operation, including equipment operating status, power generation output, load demand, meteorological conditions, and source-load dispatch. We employ techniques such as denoising, outlier detection and removal, missing data filling, and standardized conversion to improve data quality and consistency, ensuring that all subsequent analyses are based on a highly reliable dataset.

[0134] S1-3: Key feature extraction and data preprocessing

[0135] After completing basic data cleaning, we extract key characteristic variables from multi-source data, such as historical load curves, wind and solar power generation forecasts, load fluctuation indicators, and source-load matching, based on the needs of grid dispatch optimization. This extracted characteristic data is further normalized and dimensioned to eliminate scale differences between different data sources and enhance the data's applicability and stability in subsequent modeling, training, and inference stages.

[0136] S2: Markov decision process modeling, based on the standardized graph structure, defines the state space, action space, state transition probability, reward function (grid load balance, dispatch cost and system stability, etc.) and discount factor, and constructs a Markov decision framework for grid dispatch. Specifically, based on the normalized graph structure, the grid dispatch problem is formalized as a Markov decision process MDP. Specifically, it includes defining the state space, action space, and designing a reward function system based on indicators such as grid load balance, dispatch cost and system stability. This step provides a mathematical framework for the reinforcement learning model, enabling the model to accurately capture the state transition characteristics and decision logic of the grid.

[0137] The step S2 specifically includes the following steps (such as Figure 2 shown):

[0138] In the present invention, the power grid dispatch problem is modeled as a Markov decision process (MDP), and the following five-tuple is defined:

[0139]

[0140] S2-1: Define the state space

[0141] The state space describes the operating state of the power grid at a specific moment. It is defined based on the grid's operational characteristics. It encompasses various grid variables, such as wind power generation, photovoltaic power generation, energy storage device status, user load demand, grid frequency, and voltage. Each state represents the grid's operating state at a specific moment. By properly designing state variables, the model accurately captures the system's dynamic characteristics.

[0142] The state space formula is:

[0143] P t : Real-time power generation vector of each distributed power generation equipment (such as wind power and photovoltaic power).

[0144] Q t : The current charge and discharge state vector of the energy storage device.

[0145] D t : Real-time load demand vector of each user.

[0146] E t : The current topology of the power grid and the power flow status between nodes.

[0147] S2-2: Defining the action space

[0148] The action space is used to describe scheduling decisions. The action space is defined to represent the scheduling decisions that can be taken in each grid state. The action space includes strategies such as equipment startup and shutdown, generator output adjustment, energy storage device charging and discharging decisions, and load-side response control. Each action corresponds to a possible scheduling solution, which is used to guide the grid state toward the desired goal.

[0149] The action space formula is:

[0150] C t : The charging and discharging control strategy vector of the energy storage device (such as charging power, discharging power).

[0151] F t : Load demand adjustment strategy vector (such as peak shaving and valley filling of demand-side response).

[0152] S2-3: Define state transition probability (P)

[0153] Based on the definition of state and action, the state transition probability is described. That is, under the action of a specific action, the probability distribution of the power grid transitioning from the current state to the next state is described. t When performing action a t Then transfer to the next state s t+1 Considering factors such as the uncertainty of renewable energy output, load fluctuations, and equipment operation constraints, estimating the probability of the next state will help the training of reinforcement learning algorithms and the exploration of state space.

[0154] The state transition formula is: P(s t+1 |s t ,a t )

[0155] s t+1 It is determined by both grid dynamics (such as load fluctuations and power flow distribution) and control behaviors (such as energy storage scheduling).

[0156] It can be estimated by simulating the grid dynamics equations (such as power flow calculation).

[0157] S2-4: Define the reward function (R)

[0158] A reward function is designed to evaluate strategies based on the effectiveness of grid dispatch. This reward function comprehensively considers factors such as energy efficiency, power loss, minimization of dispatch costs, grid load balancing, increased renewable energy absorption, and system operational stability. This guides the model through the learning process to gradually develop a dispatch strategy that meets the overall optimization objective. The optimization goal is to maximize grid stability and efficiency while avoiding over-dispatching or under-dispatching.

[0159] The design of the reward function includes the following aspects:

[0160] Grid balancing objective: to ensure the balance between power generation and load, defined as R balance =-|P t -D t |

[0161] Scheduling cost objective: Minimize the cost of power generation and the cost of charging and discharging energy storage equipment, defined as R cost =-∑ i (c i P i,t +d i Q i,t ), where c i and d i is the cost coefficient of power generation and energy storage.

[0162] Network stability goal: reduce power flow fluctuations and line losses, defined as R stability =-∑ j∈lines L j , where L j is the power loss of the jth line.

[0163] Total reward function: R(s t ,a t )=α1R balance +α2R cost +α3R stability , where α1, α2, α3 are weight parameters.

[0164] S2-5: Define the discount factor (γ)

[0165] Set an appropriate discount factor (γ) to balance immediate rewards with long-term returns. By properly setting the discount factor, the model is encouraged to not only focus on current scheduling benefits but also consider the overall optimization effect of future system operations, enhancing the strategy's foresight and robustness.

[0166] The above processes S2-1 to S2-5 describe the power dispatch problem as a Markov decision process, and its solution goal is to find the optimal strategy π through reinforcement learning method. * , so that the cumulative reward is maximized:

[0167]

[0168] S3: Reinforcement Learning Strategy Generation. This involves interactive training using a reinforcement learning algorithm within the Markov decision framework to generate a preliminary scheduling strategy that satisfies power balance constraints. Specifically, scheduling strategy training is performed based on the defined MDP model. Through interactive learning, the model continuously adjusts its decision-making strategy based on the current state and environmental feedback in a simulated or real power grid environment to maximize cumulative rewards. During training, the model not only learns static strategies but also dynamically adapts to varying load fluctuations and renewable energy output uncertainties, forming an intelligent scheduling system with autonomous decision-making capabilities.

[0169] The step S3 specifically includes the following steps:

[0170] S3-1: Model initialization

[0171] Based on the aforementioned definition of the Markov decision process, the reinforcement learning model is initialized. Specifically, this involves constructing a fixed state space, action space, and reward function; using a parameterized neural network as an approximator for the policy function and value function; and initializing the neural network parameters. The neural network uses a dual-network structure, corresponding to the policy network and the value network, with the ReLU function as the activation function and the Adam optimizer as the optimizer.

[0172] S3-2: Environment Interaction and Strategy Learning

[0173] Build a power grid simulation environment, simulate the power grid operation process based on the power flow calculation method, and dynamically generate time series data of clean energy power generation, user load demand, and energy storage device charging and discharging behavior. At each discrete moment t, the reinforcement learning model is based on the current state s t And the policy network output π, determine the unique action a t After executing this action, the power grid environment is updated to the next state s according to the power flow calculation. t+1 , and returns the immediate reward R t .

[0174] During the interaction process, the Proximal Policy Optimization (PPO) algorithm is used to synchronously update the policy network and the value network; at the same time, the Experience Replay mechanism is introduced to perform batch sampling training on historical interaction samples to improve the independence of training data and the generalization ability of the strategy.

[0175] S3-3: Strategy Evaluation and Hyperparameter Optimization

[0176] During the reinforcement learning training process, the scheduling strategy output by the model is evaluated at a fixed period. The evaluation indicators include: the balance error between total power generation and load demand, the trend of cumulative reward changes, and the stability of the power grid flow (including voltage offset and frequency fluctuation amplitude).

[0177] Based on the evaluation results, the training hyperparameters are dynamically adjusted, including: updating the exploration rate ∈ according to the preset linear decay strategy, and adaptively adjusting the PPO clipping coefficient according to the reward stability to ensure the stability and convergence of the strategy optimization process. The update formula is:

[0178] ∈=max(∈ min ,∈ max ·e -k·t )

[0179] Among them, ∈ min is the minimum exploration rate, and k is the decay factor.

[0180] S3-4: Preliminary scheduling strategy generation

[0181] After reinforcement learning training reaches a preset stopping criterion (e.g., cumulative rewards reach a set threshold and reward fluctuations remain within a set range over multiple consecutive evaluation cycles), the final trained policy network is extracted as the preliminary scheduling strategy. This scheduling strategy, based on real-time grid state input, outputs a unique set of scheduling instructions that prioritizes clean energy generation, improves energy storage device utilization, and minimizes total grid scheduling costs while meeting load demand. This set of scheduling instructions serves as the initial solution set for the subsequent genetic algorithm optimization phase.

[0182] S4: Genetic algorithm global optimization: Using the initial scheduling strategy as the initial population, a globally optimized scheduling solution is generated through fitness function evaluation and multi-mode crossover and mutation operations. To overcome the potential local optimality issues of the reinforcement learning model, a genetic algorithm is introduced based on the initially obtained scheduling strategy for further global optimization. Through genetic operations such as population initialization, fitness evaluation, crossover and mutation, a broader space of scheduling solutions is explored, the global optimality and robustness of the scheduling strategy in complex environments are improved, and the optimal efficiency of power resource allocation and the stability of system operation are ensured.

[0183] The step S4 specifically includes the following steps:

[0184] In order to further optimize the dispatching strategy generated based on reinforcement learning, the present invention introduces a genetic algorithm to improve the efficiency and stability of power grid dispatching through global search and evolutionary optimization methods.

[0185] S4-1: Generate initial population based on preliminary scheduling plan

[0186] (1) Definition of population individuals

[0187] Each population individual represents a complete set of grid dispatch plans, including the specific power output instructions for each power generation device and energy storage device at each dispatch time. The individual encoding adopts real number encoding, with a fixed encoding structure and a length proportional to the number of dispatchable devices in the grid and the dispatch duration. The population individual formula is defined as:

[0188] x i ={G i ,C i ,F i}

[0189] in:

[0190] G i : Output adjustment strategy of power generation equipment;

[0191] C i : Charging and discharging power distribution of energy storage equipment;

[0192] F i : Load side response adjustment strategy.

[0193] (2) Generate the initial population

[0194] The preliminary scheduling strategy π generated by reinforcement learning RL A small random perturbation is introduced to generate the initial population, ensuring that the performance of the individuals in the initial population is close to the local optimum while also maintaining a certain degree of diversity. The generated solutions are also feasibility-checked to ensure they meet grid operation constraints (this invention includes checks on power balance constraints, equipment capacity limits, and other constraints). Individuals that do not meet these constraints are corrected or regenerated.

[0195]

[0196] Where N represents the population size, and the population size is fixed at N = 100 individuals.

[0197] S4-2: Fitness evaluation and individual selection

[0198] Define the fitness function f(x i ), taking the overall performance of the grid dispatching scheme as the basis for fitness evaluation, and then evaluating the pros and cons of each dispatching scheme:

[0199] f(x i )=-(λ1·L balance +θ2·L stability +λ3·C cost )

[0200] L balance : Grid power balance error, which measures the degree of matching between power input and output;

[0201] L stability : Grid operation stability indicators, such as power flow fluctuations, voltage deviation, etc.;

[0202] C cost : The operating cost of the dispatch scheme, including energy storage losses and load adjustment costs;

[0203] λ1,λ2,λ3: weight coefficients, reflecting the relative importance of different optimization objectives.

[0204] Using an elite retention strategy combined with a roulette wheel selection mechanism, we select outstanding individuals from the current population to enter the next generation based on the fitness value calculated by the fitness function, ensuring that excellent genes are passed on. The probability of an individual being selected is positively correlated with its fitness value.

[0205] S4-3: Genetic Manipulation and Evolution

[0206] (1) Crossover operation

[0207] Perform a crossover operation on the selected individuals to generate the next generation:

[0208] x new =α·x i +(1-α)·x j

[0209] where x i ,x j represents the parent individuals, and α is the crossover ratio, which is usually a random value.

[0210] For the crossover operation of the grid dispatching scheme, three crossover schemes are adopted. First, the generation dispatching part G in the dispatching scheme is directly i 、Energy storage dispatch part C i and demand response part F i Each of these three gene segments is considered as a gene fragment, and two parent individuals are randomly selected to cross the above three gene segments. Since power grid scheduling usually has time series characteristics, a scheduling cycle can be divided into multiple time periods (such as morning peak, midday flat and valley, evening peak, etc.). From the two parent individuals, the scheduling strategies of different time periods are selected and combined into new individuals. Finally, since the power grid scheduling plan involves multiple regions (such as different load nodes or energy storage equipment), the scheduling strategies of parent individuals can be exchanged between different regions.

[0211] (2) Mutation operation

[0212] Perform genetic mutation on some individuals to enhance population diversity and avoid falling into local optimality:

[0213] x mut =x k +δ

[0214] Where δ is the random perturbation.

[0215] For the grid dispatching plan, the energy storage dispatching part C in the dispatching plan can be i and demand response part F i Perform random parameter perturbations to introduce larger random variations to lower-priority equipment in the power grid (such as partial load nodes or auxiliary energy storage) to maintain the scheduling stability of key equipment.

[0216] Based on crossover and mutation operations, a new generation of population is generated: The population is optimized in a loop. To ensure the quality of the solution, the solution with the highest fitness in the previous generation is directly retained in the next generation population.

[0217] S4-4: Output optimal strategy

[0218] When the population fitness value changes less than the set threshold (or the population evolves to the maximum iteration number, which is fixed at 200 generations in this invention), the genetic algorithm converges. Finally, the individual with the highest fitness is selected from the population, and the grid dispatching scheme corresponding to its code is extracted as the optimal dispatching strategy after genetic algorithm optimization:

[0219]

[0220] S5: Adaptive feedback control deploys the globally optimized dispatching plan to the actual power grid. The policy network parameters are dynamically updated based on real-time data collected from node loads, power output, and frequency, forming a closed-loop control process of "data collection - policy update - effect evaluation." During the dispatching strategy deployment and operation phase, grid operation data is continuously collected in real time to establish a feedback loop. Based on a reinforcement learning model, the dispatching strategy is dynamically updated to rapidly adapt to factors such as changes in grid structure, load fluctuations, and fluctuations in renewable energy output. Through continuous iterative optimization, the dispatching strategy achieves adaptive evolution, further enhancing the intelligence, stability, and reliability of power grid operations.

[0221] The step S5 specifically includes the following steps (such as Figure 3 shown):

[0222] S5-1: Real-time feedback data collection and preprocessing

[0223] During the grid dispatch process under the guidance of the global optimization dispatching scheme obtained by S4 step optimization, real-time grid operation data is collected, including but not limited to: node load demand L i(t) , node power generation output G i(t) , Energy storage device status (remaining capacity S i(t) ), power grid system frequency f(t), line power transmission status P {ij}(t) Where i and j represent the node numbers in the network, and t is the current time.

[0224] The collected data is normalized (normalized to the [0,1] interval) and anomaly detection (Z-score method) to eliminate invalid or abnormal data to ensure the accuracy of subsequent feedback learning.

[0225] S5-2: Feedback Learning and Incremental Update of Reinforcement Learning Model

[0226] Based on the real-time preprocessed data, the reinforcement learning model performs incremental updates. Specifically, the model learns according to the following rules:

[0227] Current status t It is composed of the collected real-time grid operation data;

[0228] Action a t Determined by the current scheduling policy;

[0229] Reward t Calculated based on actual grid operation results.

[0230] S5-3: Real-time evaluation and adaptive adjustment of strategy effects

[0231] During the feedback learning process, the updated scheduling strategy is evaluated periodically (every hour or every 100 interactions). If the evaluation indicator exceeds the set threshold (load satisfaction rate is less than 95%, system frequency deviation exceeds ±0.1Hz, etc.), the adaptive strategy adjustment mechanism is triggered:

[0232] Temporarily increase the exploration action ratio;

[0233] Update the reward function parameter w i ;

[0234] Re-adjust the output distribution of the strategy network to enhance the model's adaptability to abnormal working conditions.

[0235] S5-4: Forming a continuous optimization closed loop

[0236] By continuously executing steps S5-1 through S5-3, a closed-loop mechanism of data collection, feedback learning, strategy evaluation, and adaptive adjustment is formed. This enables the grid dispatch strategy to dynamically adapt to external environmental disturbances such as load fluctuations, climate change, and changes in equipment status, continuously improving the stability, reliability, and economic efficiency of grid operation. The strategy optimized by the genetic algorithm is also compared with the strategy initially generated by reinforcement learning to verify the optimization effect of the genetic algorithm.

[0237] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in the present invention should be included in the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection of the claims.

Claims

1. A smart grid dispatch optimization method based on adaptive evolutionary control, characterized in that: The steps include: S1: Power network graph modeling and standardization. Wind power / photovoltaic equipment, energy storage equipment, and user loads in the power network are modeled as graph nodes. Edge connections are established between graph nodes through power transmission paths. Various dynamic data generated during the operation of the power grid are denoised, gap-filled, and normalized to generate a standardized graph structure containing device characteristic parameters. S2: Markov decision process modeling, based on the standardized graph structure, defines the state space, action space, state transition probability, reward function and discount factor, and constructs a Markov decision framework for power grid scheduling; S3: Reinforcement learning strategy generation, using the reinforcement learning algorithm to perform interactive training under the Markov decision framework to generate a preliminary scheduling strategy that meets the power balance constraint; S4: Genetic algorithm global optimization, using the preliminary scheduling strategy as the initial population, generating a globally optimized scheduling solution through fitness function evaluation and multi-mode crossover and mutation operations; S5: Adaptive feedback control deploys the globally optimized dispatching scheme to the actual power grid and dynamically updates the strategic network parameters based on real-time collected node load, power output and frequency data.

2. The smart grid dispatch optimization method based on adaptive evolutionary control according to claim 1, characterized in that: The step S1 specifically includes the following steps: S1-1: Modeling of power network equipment; Wind power / photovoltaic equipment, energy storage equipment, and user loads are abstracted as nodes in a graph, and the key characteristics of the equipment are described as node attributes. Device nodes are connected based on power flow relationships to establish graph edges. The key characteristics include power generation capacity, storage capacity, and load variation characteristics. Edge attributes include transmission line capacity, power transmission capacity, and energy loss coefficient. S1-2: Power system operation data collection and cleaning; Based on the S1-1 modeling framework, various dynamic data generated during the operation of the power grid are systematically collected. De-noising, outlier detection and elimination, missing data filling, and standardized conversion are applied to various dynamic data. S1-3: Key feature extraction and data preprocessing; After completing the basic data cleaning, key characteristic variables are extracted from multi-source data based on the needs of power grid dispatch optimization, including historical load curves, wind and solar power generation forecasts, load fluctuation indicators, and source-load matching. The extracted key characteristic variables are normalized and dimensionalized to eliminate scale differences between different data sources and enhance the applicability and stability of the data in subsequent modeling, training, and inference stages.

3. The smart grid dispatch optimization method based on adaptive evolutionary control according to claim 2, characterized in that: The various types of dynamic data include equipment operating status, power generation output, load demand, meteorological conditions, and source-load scheduling direction.

4. The smart grid dispatch optimization method based on adaptive evolutionary control according to claim 1, characterized in that: The step S2 specifically includes the following steps: The grid dispatch problem is modeled as a Markov decision process, which is defined as the following five-tuple: S2-1: Define the state space The state space is used to describe the operating state of the power grid at a certain moment. It is defined based on the operating characteristics of the power grid. The state space contains various variables of the power grid, including wind power generation, photovoltaic power generation, the status of energy storage equipment, user load demand, grid frequency, and voltage. Each state represents the operating state of the power grid at a certain moment. The state space formula is: P t : Real-time power generation vector of each distributed power generation device; Q t : The current charge and discharge state vector of the energy storage device; D t : Real-time load demand vector of each user; E t : The current topology of the power grid and the power flow status between nodes; S2-2: Defining the action space The action space is used to describe scheduling decisions. The action space is defined to represent the scheduling decisions that the grid can take in each state. The action space includes equipment startup and shutdown, generator output adjustment, energy storage device charging and discharging decisions, and load-side response control. Each action corresponds to a possible scheduling solution, which is used to guide the grid state toward the desired goal. The action space formula is: C t : charge and discharge control strategy vector of energy storage equipment; F t : Load demand adjustment strategy vector; S2-3: Define state transition probability (P); Based on the definition of state and action, the state transition probability is characterized, that is, the probability distribution of the power grid transferring from the current state to the next state under the action of a specific action, and the probability distribution of the power grid transferring from the current state s t When performing action a t Then transfer to the next state s t+1 possibility; The state transition formula is: P(s t+1 |s t ,a t ); S t+1 It is determined by the grid dynamics and control behavior and is estimated by simulating the grid dynamics equations; S2-4: Define the reward function (R); Design a reward function to evaluate strategies based on the effectiveness of grid dispatch. The reward function comprehensively considers energy efficiency, power loss, minimization of dispatch costs, grid load balance, improvement of renewable energy absorption rate, and system operation stability, thereby guiding the model to gradually form a dispatch strategy that meets the overall optimization goal during the learning process. The design of the reward function includes: Grid balancing objective: to ensure the balance between power generation and load, defined as R balance =-|P t -D t |; Scheduling cost objective: Minimize the cost of power generation and the cost of charging and discharging energy storage equipment, defined as R cost =-∑ i (c i P i,t +d i Q i,t ), where c i and d i is the cost coefficient of power generation and energy storage; Network stability goal: reduce power flow fluctuations and line losses, defined as R stability =-∑ j∈lines L j , where L j is the power loss of the jth line; Total reward function: R(s t ,a t )=α1R balance +α2R cost +α3R stability , where α1, α2, α3 are weight parameters; S2-5: Define the discount factor (γ); Set an appropriate discount factor to balance immediate rewards with long-term returns. By properly setting the discount factor value, the model is encouraged to focus not only on current scheduling benefits but also on the overall optimization effect of future system operations, thereby enhancing the strategy's foresight and robustness. Steps S2-1 to S2-5 describe the power dispatch problem as a Markov decision process, whose solution goal is to find the optimal strategy π through reinforcement learning method. * , so that the cumulative reward is maximized:

5. The smart grid dispatch optimization method based on adaptive evolutionary control according to claim 1, characterized in that: The step S3 specifically includes the following steps: S3-1: Model initialization; Based on the definition of the Markov decision process, the reinforcement learning model is initialized, specifically including: constructing a fixed state space, action space, and reward function; using a parameterized neural network as an approximator of the policy function and the value function respectively; initializing the parameters of the neural network; the neural network adopts a dual network structure, corresponding to the policy network and the value network respectively, using the ReLU function as the activation function, and using the Adam optimizer as the optimizer; S3-2: Environmental Interaction and Strategy Learning; Construct a power grid simulation environment, simulate the power grid operation process based on the power flow calculation method, and dynamically generate time series data of clean energy power generation, user load demand, and energy storage device charging and discharging behavior; at each discrete moment t, the reinforcement learning model is used according to the current state s t And the policy network output π, determine the unique action a t After executing this action, the power grid environment is updated to the next state s according to the flow calculation t+1 , and returns the immediate reward R t ; During the interaction process, the proximal policy optimization algorithm is used to synchronously update the policy network and the value network. At the same time, the experience replay mechanism is introduced to perform batch sampling training on historical interaction samples to improve the independence of training data and the generalization ability of the strategy. S3-3: Strategy evaluation and hyperparameter optimization; During the reinforcement learning training process, the dispatching strategy output by the model is evaluated at a fixed period. The evaluation indicators include: the balance error between total power generation and load demand, the trend of cumulative reward changes, and the stability of the power flow of the power grid, which includes voltage offset and frequency fluctuation amplitude. Based on the evaluation results, the training hyperparameters are dynamically adjusted, including updating the exploration rate ∈ according to the preset linear decay strategy and adaptively adjusting the PPO clipping coefficient according to the reward stability to ensure the stability and convergence of the strategy optimization process. The update formula is: ∈=max(∈ min ,∈ max ·and -k·t ); Among them, ∈ min is the minimum exploration rate, k is the decay factor; S3-4: Preliminary scheduling strategy generation; After the reinforcement learning training reaches the preset stopping criteria, the final trained policy network is extracted as the preliminary scheduling strategy. The stopping criteria are that the cumulative reward reaches the set threshold and the reward fluctuation is less than the set range for multiple consecutive evaluation cycles.

6. The smart grid dispatch optimization method based on adaptive evolutionary control according to claim 1, characterized in that: The step S4 specifically includes the following steps: S4-1: Generate initial population based on preliminary scheduling plan (1) Definition of individuals in a population; Each population individual represents a complete set of grid dispatch plans, including the specific power output instructions of each power generation equipment and energy storage equipment at each dispatch time. The individual encoding adopts real number encoding, with a fixed encoding structure and a length proportional to the number of dispatchable devices in the grid and the dispatch duration. The population individual formula is defined as: x i ={G i ,C i ,F i }; in: G i : Output adjustment strategy of power generation equipment; C i : Charging and discharging power distribution of energy storage equipment; F i : Load side response adjustment strategy; (2) Generate the initial population; The preliminary scheduling strategy π generated by reinforcement learning RL , introducing a small random perturbation to generate the initial population, ensuring that the performance of the individuals in the initial population is close to the local optimum and has a certain degree of diversity; at the same time, the feasibility of the generated scheme is checked to ensure that it meets the grid operation constraints. Individuals that do not meet the constraints will be corrected or regenerated; Where N represents the population size, which is fixed at N = 100 individuals; S4-2: Fitness evaluation and individual selection; Define the fitness function f(x i ), taking the overall performance of the grid dispatching scheme as the basis for fitness evaluation, and then evaluating the pros and cons of each dispatching scheme: f(x i )=-(λ1·L balance +λ2·L stability +λ3·C cost ); L balance : Grid power balance error, which measures the degree of matching between power input and output; L stability : Grid operation stability indicators, such as power flow fluctuations, voltage deviation, etc.; C cost : The operating cost of the dispatch scheme, including energy storage losses and load adjustment costs; λ1,λ2,λ3: weight coefficients, reflecting the relative importance of different optimization objectives; An elite retention strategy combined with a roulette wheel selection mechanism is used to select outstanding individuals from the current population to enter the next generation based on the fitness value calculated by the fitness function, ensuring that excellent genes are passed on. The probability of an individual being selected is positively correlated with its fitness value. S4-3: Genetic manipulation and evolution; (1) Crossover operation; Perform a crossover operation on the selected individuals to generate the next generation: x new =α·x i +(1-a)·x j ; where x i ,x j represents the parent individual, α is the crossover ratio, which is a random value; For the cross operation of the grid dispatching scheme, three cross schemes are adopted: First, the power generation dispatching part G in the dispatching scheme is directly i 、Energy storage dispatch part C i and demand response part F i Each of these three gene fragments is considered a gene fragment, and two parent individuals are randomly selected to cross the three gene fragments. Since power grid scheduling usually has time series characteristics, a scheduling cycle is divided into multiple time periods. From the two parent individuals, the scheduling strategies of different time periods are selected and combined into new individuals. Finally, since the power grid scheduling plan involves multiple regions, the scheduling strategies of the parent individuals are exchanged between different regions. (2) mutation operation; Perform genetic mutations on some individuals to enhance population diversity and avoid falling into local optimality: x mut =x k +δ; Where δ is the random perturbation; For the grid dispatching plan, the energy storage dispatching part C in the dispatching plan can be i and demand response part F i Perform random parameter perturbations to introduce larger random variations to lower-priority devices in the power grid, maintaining the dispatch stability of key devices. Based on crossover and mutation operations, a new generation of population is generated: The population is optimized step by step through a loop. To ensure the quality of the solution, the solution with the highest fitness in the previous generation is directly retained in the population of the next generation. S4-4: Output the optimal strategy; When the change in the population fitness value is less than the set threshold, the genetic algorithm converges; eventually, the individual with the highest fitness is selected from the population, and the grid dispatching scheme corresponding to its code is extracted as the optimal dispatching strategy after genetic algorithm optimization:

7. The smart grid dispatch optimization method based on adaptive evolutionary control according to claim 1, characterized in that: The step S5 specifically includes the following steps: S5-1: Real-time feedback data collection and preprocessing; During the grid dispatch process under the guidance of the global optimization dispatch scheme obtained by S4 step optimization, grid operation data is collected in real time, including: node load demand L i(t) , node power output G i(t) , remaining capacity of energy storage equipment S i(t) , grid system frequency f(t), line power transmission status P {ij}(t) ; Where i, j represent the node numbers in the network, and t is the current time; The collected data is processed through standardization and anomaly detection to eliminate invalid or abnormal data; S5-2: Feedback learning and incremental updating of reinforcement learning models; Based on the real-time preprocessed data, the reinforcement learning model performs incremental updates. Specifically, the model learns according to the following rules: Current status t It is composed of the collected real-time grid operation data; Action a t Determined by the current scheduling policy; Reward t Calculated based on actual grid operation results; S5-3: Real-time evaluation and adaptive adjustment of strategy effects; During the feedback learning process, the updated scheduling policy is periodically evaluated. If the evaluation indicator exceeds the set threshold, the adaptive policy adjustment mechanism is triggered: Temporarily increase the exploration action ratio; Update the reward function parameter w i ; Re-adjust the output distribution of the strategy network to enhance the model's adaptability to abnormal working conditions; S5-4: Form a continuous optimization closed loop; Through the continuous execution of steps S5-1 to S5-3, a closed-loop mechanism of data collection - feedback learning - strategy evaluation - adaptive adjustment is formed, so that the power grid dispatching strategy can dynamically adapt to external environmental disturbances such as load fluctuations, climate change and equipment status changes, and continuously improve the stability, reliability and economy of power grid operation; at the same time, the strategy optimized by the genetic algorithm is compared with the strategy initially generated by reinforcement learning to verify the optimization effect of the genetic algorithm.

8. The smart grid dispatch optimization method based on adaptive evolutionary control according to claim 7, characterized in that: The thresholds set in step S5-3 are that the load satisfaction rate is less than 95% and the system frequency deviation exceeds ±0.1 Hz.

Images